Production Plant Control Using Uncertainty-Corrected Design Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing design optimization systems for complex products consume excessive computational resources due to inaccurate surrogate models and the need for numerous high-fidelity simulations, which are costly and inefficient.
Innovation Solution
Implement a machine learning module trained in disjoint phases using simulation results to predict performance quantities and uncertainties, with a correction factor derived from simulation data to adjust predictions, allowing simulations to be skipped when predictions are accurate or indicate poor performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If surrogate models are used to predict simulation outcomes, then computational resources are reduced, but prediction accuracy deteriorates
Solution Approach 1:
The system implements a feedback mechanism where simulation results are used to train and update the machine learning model iteratively. The model predicts outcomes, these predictions are compared with actual simulation results, and the discrepancy (loss) is used to update the model parameters. This closed-loop feedback enables the surrogate model to progressively improve its accuracy while maintaining computational efficiency.
Solution Approach 2:
The machine learning model is trained in advance using previously obtained simulation data before being deployed for prediction. This preliminary training action creates a pre-computed knowledge base that can quickly answer prediction queries without requiring full simulation execution, thus reducing computational resources while maintaining accuracy for the trained domain.
2Manufacturing precision
If numerous high-fidelity simulations are performed to optimize design, then design precision is improved, but computational cost increases
Solution Approach 1:
Instead of performing exhaustive simulations for all possible design variants, the system uses the machine learning model to evaluate and filter design options. The model performs partial evaluation by predicting outcomes for many designs quickly, allowing only the most promising candidates to undergo full high-fidelity simulation. This selective approach achieves adequate design optimization with significantly reduced computational cost.
Solution Approach 2:
The machine learning model serves as a computational copy or surrogate of the expensive simulation process. Rather than repeatedly executing the full high-fidelity simulation for each design variant, the system uses the trained model (a simplified copy) to predict outcomes. This copying approach maintains acceptable accuracy while dramatically reducing the computational energy required for design optimization.
3Productivity
If machine learning predictions are used to skip simulations, then productivity is improved, but reliability of performance assessment deteriorates
Solution Approach 1:
The system dynamically adjusts the decision to run or skip simulations based on the predicted uncertainty. When the model predicts high uncertainty for a design variant, the system automatically decides to perform a full simulation to ensure reliable assessment. When uncertainty is low, skipping the simulation maintains productivity. This dynamic adaptation balances reliability and productivity based on actual prediction confidence.
Solution Approach 2:
The system uses feedback from the comparison between predicted and actual simulation results to improve future decisions. When simulations are performed, the results feed back into the training data, improving the model's ability to accurately assess when predictions are reliable enough to skip simulations. This feedback loop progressively enhances both the reliability of predictions and the accuracy of the skip/execute decision.
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
A simulator (SIM) for determining simulated performance quantities (PRS) for design variants of the product (P) and a machine learning module (BNN) to be trained to determine predictive performance quantities (PRP) and their uncertainties (UC) are provided. Furthermore, a multitude of design data records (DR) each specifying a design variant are generated. A respective design data record (DR) is fed into the machine learning module (BNN), and/or the corresponding design variant is simulated. Using simulated performance quantities (PRS) as training data, the machine learning module (BNN) is trained in disjoint training phases (TP1, TP2, TP3). Between the disjoint training phases a correction factor (CF) for uncertainties (UC) determined by the machine learning module (BNN) is derived by means of simulated performance quantities (PRS). The correction factor (CF) is used to correct future uncertainties (UC). The corrected uncertainties (UC') are then used to decide whether to run or to skip a simulation. Accordingly, a performance value (PV) is derived from the simulated performance quantity (PRS) if the simulation is run, or otherwise from the predictive performance quantity (PRP) determined by the machine learning module. Depending on the generated performance values (PV), a performance-optimizing design data record (ODR) is determined from the generated design data records (DR). The performance-optimizing design data record (ODR) is then output for controlling the production plant.